First run
Start uji in the directory you want to work on.
cd ~/code/project
uji
Signing in
Type /login and pick a provider from the list.
- Anthropic and OpenAI ask how you want to sign in. “Subscription (sign in with browser)” opens the provider’s sign-in page, and “API key” asks for a key.
- OpenAI also offers “Subscription (sign in with device code)”. Open the displayed link on any device and enter the one-time code within 15 minutes. Enable device code login in your ChatGPT security settings or workspace permissions first. This works without a browser or a callback port on the machine running uji.
- A provider that needs a key asks for it. The prompt names the environment
variable uji also reads, such as
ANTHROPIC_API_KEY, and pressing Enter on an empty prompt uses that variable instead. - GitHub Copilot starts device code sign-in without a method picker. Open the displayed link and enter the one-time code. uji uses your Copilot subscription and loads its chat models with tool support. GitHub Models remains a separate provider.
- Custom asks for the
base_urland model of a server that takes OpenAI chat requests. - A provider you are signed in to asks whether to sign in again or sign out.
uji saves keys and subscription sign-ins in auth.toml in the
data directory, readable only by you. To keep them
in the system keychain instead, call
uji.auth.configure in your config.
Choosing a model
/models lists the models of every provider you are signed in to. /effort
sets how much the model reasons, from the levels that model accepts.
Making it yours
/setup writes a starter config and asks what you want in it.
Starting a config shows what it
writes.
Working with the model
- Enter sends your message. Shift+Enter, Alt+Enter or Ctrl+J start a new line.
- A message you send while the model works waits, and goes out when the turn ends.
- Esc clears the input line. On an empty line, Esc stops the turn.
- A line that starts with
!runs in your shell, in the session’s directory. Its output shows in the transcript, and the model does not see it. - When a tool call needs your approval,
ylets it run.nor Esc refuses it, and you can then tell the model what to do instead.
Sending images
Ctrl+V attaches the image on the clipboard, or pastes its text when it holds
no image. Pasting the path of an image file attaches that file, which is what
dragging a file into most terminals does, and @screenshot.png in a message
attaches the file when you send it.
Each attached image shows as [image #1] on the input line, and deleting the
marker drops the image. uji takes PNG, JPEG, GIF and WebP files. It turns
photos upright and scales images down to 1568 pixels on their long side, then
to a smaller size or JPEG when they would still pass 5 MB.
A request carries the 20 newest images of the conversation, and a note takes the place of older ones. A model that takes no images gets the text with the same kind of note. uji knows this for most built-in models, and learns it for others the first time a model refuses a request with images.
The model can open images on its own as well. read_file on a PNG, JPEG, GIF
or WebP file in the working directory gives it the image.
uji saves every conversation as a session. Commands shows how to go back to one.